def __init__(self, args):
super(GFootballEnv, self).__init__()
self.representation = args.state.split('_')[0]
self.stacked = 'stacked' in args.state
# wrap the original environment
# https://github.com/google-research/football/blob/master/gfootball/env/__init__.py
# this is same as online environment
self.raw_env = football_env.create_environment(
env_name=args.level,
stacked=False, # set to False to align with online evaluation
representation="raw", # set to raw to align with online evaluation
rewards=args.reward_experiment)
self.obs_stack = deque([], maxlen=4)
# this is for training environment
self.real_env = football_env.create_environment(
env_name=args.level,
stacked=self.
stacked, # set to False to align with online evaluation
representation=self.representation,
rewards=args.reward_experiment)
self.action_space = self.real_env.action_space
self.observation_space = self.real_env.observation_space
def random():
# render the environment
display_train = False
if display_train:
env = football_env.create_environment(env_name='academy_empty_goal',
representation='pixels',
render=True)
else:
env = football_env.create_environment(env_name='academy_empty_goal',
representation='simple115',
render=True)
total_rewards = 0.0
total_steps = 0
obs = env.reset()
# start an episode.
while True:
action = env.action_space.sample()
# choose a random action
obs, reward, done, _ = env.step(action)
# obtain rewards
total_rewards += reward
total_steps += 1
if done:
break
print("episode done %d steps. total rewards %.2f" %
(total_steps, total_rewards))
def __init__(self, args):
self.num_agents = args.num_agents
self.scenario_name = args.scenario_name
# make env
if not (args.use_render and args.save_videos):
self.env = football_env.create_environment(
env_name=args.scenario_name,
stacked=args.use_stacked_frames,
representation=args.representation,
rewards=args.rewards,
number_of_left_players_agent_controls=args.num_agents,
number_of_right_players_agent_controls=0,
channel_dimensions=(args.smm_width, args.smm_height),
render=(args.use_render and args.save_gifs))
else:
# render env and save videos
self.env = football_env.create_environment(
env_name=args.scenario_name,
stacked=args.use_stacked_frames,
representation=args.representation,
rewards=args.rewards,
number_of_left_players_agent_controls=args.num_agents,
number_of_right_players_agent_controls=0,
channel_dimensions=(args.smm_width, args.smm_height),
# video related params
write_full_episode_dumps=True,
render=True,
write_video=True,
dump_frequency=1,
logdir=args.video_dir)
self.max_steps = self.env.unwrapped.observation()[0]["steps_left"]
self.remove_redundancy = args.remove_redundancy
self.zero_feature = args.zero_feature
self.share_reward = args.share_reward
self.action_space = []
self.observation_space = []
self.share_observation_space = []
if self.num_agents == 1:
self.action_space.append(self.env.action_space)
self.observation_space.append(self.env.observation_space)
self.share_observation_space.append(self.env.observation_space)
else:
for idx in range(self.num_agents):
self.action_space.append(
spaces.Discrete(n=self.env.action_space[idx].n))
self.observation_space.append(
spaces.Box(low=self.env.observation_space.low[idx],
high=self.env.observation_space.high[idx],
shape=self.env.observation_space.shape[1:],
dtype=self.env.observation_space.dtype))
self.share_observation_space.append(
spaces.Box(low=self.env.observation_space.low[idx],
high=self.env.observation_space.high[idx],
shape=self.env.observation_space.shape[1:],
dtype=self.env.observation_space.dtype))
def main():
display_train = False
# render environment
if display_train:
env = football_env.create_environment(env_name='academy_empty_goal',
representation='pixels',
render=True)
else:
env = football_env.create_environment(env_name='academy_empty_goal',
representation='simple115',
render=True)
print("action space=", env.action_space)
print("observation space", env.observation_space)
def __init__(self,
env_name='academy_empty_goal_close',
representation='simple115',
rewards='scoring,checkpoints',
stacked=False,
dump_path=None,
solved_at=1,
**kwargs):
if football_env is None:
raise ImportError('Could not import gfootball! '
'HINT: Perform the setup instructions here: '
'https://github.com/google-research/football')
self._solved_at = solved_at
self._env = football_env.create_environment(
env_name=env_name,
representation=representation,
rewards=rewards,
stacked=stacked,
write_full_episode_dumps=dump_path is not None,
write_goal_dumps=False,
logdir=dump_path or '',
**kwargs)
self.action_space = self._env.action_space
self.observation_space = self._env.observation_space
def watch(self, *, env, matches, weights, record):
environment = SubprocVecEnv([
lambda: football.create_environment(
env_name = "11_vs_11_easy_stochastic",
representation = self.configs[0]["representation"],
rewards = self.configs[0]["rewards"],
enable_goal_videos = False,
enable_full_episode_videos = True,
render = True,
write_video = record,
dump_frequency = 1,
logdir = "/home/charlie/Projects/Python/Football/videos/",
extra_players = self.configs[0]["extra_players"],
number_of_left_players_agent_controls = self.configs[0]["number_of_left_players_agent_controls"],
number_of_right_players_agent_controls = self.configs[0]["number_of_right_players_agent_controls"],
enable_sides_swap = self.configs[0]["enable_sides_swap"]
) for _ in range(1)
])
# self.model.set_env(environment)
watch = PPO2.load(weights, env = environment)
for match in range(matches):
watch.learn(total_timesteps = 3100)
def create_gfootball_multiagent_env(save_dir, dump_freq, render):
num_agents = 3 # for 'academy_3_vs_1_with_keeper' scenario
env = football_env.create_environment(
env_name='academy_3_vs_1_with_keeper',
stacked=False,
representation='simple115',
rewards='scoring',
logdir=str(save_dir/'football_dumps'),
enable_goal_videos=False,
enable_full_episode_videos=bool(dump_freq),
render=render,
write_video=True,
dump_frequency=dump_freq,
number_of_left_players_agent_controls=3,
number_of_right_players_agent_controls=0,
enable_sides_swap=False,
channel_dimensions=(
observation_preprocessing.SMM_WIDTH,
observation_preprocessing.SMM_HEIGHT)
)
action_spaces = [gym.spaces.Discrete(env.action_space.nvec[1]) for _ in range(num_agents)]
observation_spaces = [gym.spaces.Box(
low=env.observation_space.low[0],
high=env.observation_space.high[0],
dtype=env.observation_space.dtype) for _ in range(num_agents)]
return env, action_spaces, observation_spaces
def get_env_info(env_name, state, reward_experiment, num_left_agents,
num_right_agents, representation, render, seed, num_agents):
is_football = '11' in env_name or 'academy' in env_name
if is_football:
env = football_env.create_environment(
representation=representation,
env_name=env_name,
stacked=('stacked' in state),
rewards=reward_experiment,
logdir=None,
render=render and (seed == 0),
dump_frequency=50 if render and seed == 0 else 0)
else:
env = create_atari_mjc_env(env_name)
if num_agents == 1:
from a2c_ppo_acktr.envs import ObsUnsqueezeWrapper
env = ObsUnsqueezeWrapper(env)
env.reset()
num_left_player = env.unwrapped._cached_observation[0]['left_team'].shape[
0] if is_football else 1
num_right_player = env.unwrapped._cached_observation[0][
'right_team'].shape[0] if is_football else 0
observation_space = copy.deepcopy(env.observation_space)
action_space = copy.deepcopy(env.action_space)
env.close()
return num_left_player, num_right_player, observation_space, action_space
def _create_single_football_env(
level,
stacked,
representation,
reward_experiment,
write_goal_dumps,
write_full_episode_dumps,
write_video,
dump_frequency,
render,
process_number=0,
):
"""
Creates gfootball environment.
Meaning of all variables you can find in footbal.gfootball.examples.run_ppo2.py
"""
env = create_environment(
env_name=level,
stacked=stacked,
representation=representation,
rewards=reward_experiment,
logdir=logger.get_dir(),
write_goal_dumps=write_goal_dumps and (process_number == 0),
write_full_episode_dumps=write_full_episode_dumps
and (process_number == 0),
write_video=write_video,
render=render and (process_number == 0),
dump_frequency=dump_frequency if render and process_number == 0 else 0)
env = monitor.Monitor(
env,
logger.get_dir()
and os.path.join(logger.get_dir(), str(process_number)))
return env
def main(args):
print(args.early_stop)
print(args.disable_actions)
set_seed(args.seed)
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
# initialize environment
n_players = 3
env = football_env.create_environment(env_name="academy_3_vs_1_with_keeper",
representation="simple115",
number_of_left_players_agent_controls=n_players,
stacked=False,
logdir="/tmp/football",
write_goal_dumps=False,
write_full_episode_dumps=False,
render=False)
# state and action space
state_space_size = env.observation_space.shape[1] # we are using simple115 representation
action_space_size = env.action_space.nvec.tolist()[0] # 三个 players 动作空间相同
action_list = list(range(action_space_size))
# state[98:100] 表示控制的三个球员
if args.disable_actions:
action_list = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]
action_space_size = len(action_list)
# model
print("loading models")
actors = [Actor(state_space_size=state_space_size, action_space_size=action_space_size) for _ in range(n_players)]
critic = Critic(state_space_size=state_space_size)
# maa2c
maa2c = MAA2C(args=args, env=env, actors=actors, critic=critic, action_list=action_list, device=device)
print("learn")
maa2c.learn()
def main(args):
set_seed(args.seed)
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
# initialize environment
n_players = 3
env = football_env.create_environment(
env_name="academy_3_vs_1_with_keeper",
representation="simple115",
number_of_left_players_agent_controls=n_players,
stacked=False,
logdir="/tmp/football",
write_goal_dumps=False,
write_full_episode_dumps=False,
render=False)
# state and action space
state_space_size = env.observation_space.shape[
1] # we are using simple115 representation
action_space_size = env.action_space.nvec.tolist()[0] # 三个 players 动作空间相同
# state[98:100] 表示控制的三个球员
# model
print("loading models")
actors = [
Actor(state_space_size=state_space_size,
action_space_size=action_space_size) for _ in range(n_players)
]
critics = [
Critic(state_space_size=state_space_size,
action_space_size=action_space_size,
n_players=n_players) for _ in range(n_players)
]
old_actors = [
Actor(state_space_size=state_space_size,
action_space_size=action_space_size) for _ in range(n_players)
]
old_critics = [
Critic(state_space_size=state_space_size,
action_space_size=action_space_size,
n_players=n_players) for _ in range(n_players)
]
for old_actor, actor in zip(old_actors, actors):
old_actor.load_state_dict(actor.state_dict())
for old_critic, critic in zip(old_critics, critics):
old_critic.load_state_dict(critic.state_dict())
# maddpg
maddpg = MADDPG(env=env,
action_list=list(range(action_space_size)),
actors=actors,
critics=critics,
old_actors=old_actors,
old_critics=old_critics,
args=args,
device=device)
print("learn")
maddpg.learn()
def __init__(self,
env_name="11_vs_11_easy_stochastic",
obs_representation="stacked_smm",
rewards="scoring,checkpoints",
logdir="/tmp/football",
env_id=0):
super(FootballEnvWrapper, self).__init__()
print("Env: " + env_name)
self.env = football_env.create_environment(
env_name=env_name,
stacked=False,
representation='raw',
rewards=rewards,
logdir=logdir,
write_goal_dumps=False and (env_id == 0),
write_full_episode_dumps=False and (env_id == 0),
write_video=False and (env_id == 0),
render=False,
dump_frequency=30)
self.obs_representation = obs_representation
# For Frame Stack
self.stacked_obs = collections.deque([], maxlen=4)
# Define observation space and action space
# They must be gym.spaces objects
self.action_space = gym.spaces.Discrete(19) # 19 actions
if obs_representation == "smm":
self.observation_space = gym.spaces.Box(low=0,
high=255,
shape=(72, 96, 4),
dtype=np.uint8)
elif obs_representation == "stacked_smm":
self.observation_space = gym.spaces.Box(low=0,
high=255,
shape=(72, 96, 16),
dtype=np.uint8)
elif obs_representation == "float115":
self.observation_space = gym.spaces.Box(low=-30.0,
high=30.0,
shape=(115, ),
dtype=np.float32)
elif obs_representation == "pixels":
pass
elif obs_representation == "raw":
# Use OBSParser
self.observation_space = gym.spaces.Box(low=-30.0,
high=30.0,
shape=(207, ),
dtype=np.float32)
self.ball_owned_team = -1
self.rewards = rewards
def __init__(self, seed_dir, dump_freq, representation, render):
self.nagents = 3 # only using 'academy_3_vs_1_with_keeper' level for the moment
self.representation = representation
if representation == "simple37":
env_representation = "simple115" # we convert obs from 115 to 37 observations ourselves
else:
env_representation = representation
# Instantiate environment
self.env = football_env.create_environment(
env_name='academy_3_vs_1_with_keeper',
stacked=False,
representation=env_representation,
rewards='scoring',
logdir=str(seed_dir / 'football_dumps'),
enable_goal_videos=False,
enable_full_episode_videos=bool(dump_freq),
render=render,
write_video=True,
dump_frequency=dump_freq,
number_of_left_players_agent_controls=3,
number_of_right_players_agent_controls=0,
enable_sides_swap=False,
channel_dimensions=(observation_preprocessing.SMM_WIDTH,
observation_preprocessing.SMM_HEIGHT))
obs_space_low = self.env.observation_space.low[0]
obs_space_high = self.env.observation_space.high[0]
# Adapting obs_space properties if we transform simple115 to simple37
if self.representation == "simple37":
obs_space_low = self.convert_simple115_to_simple37(
simple115_vectors=np.expand_dims(obs_space_low, axis=0))[0]
obs_space_high = self.convert_simple115_to_simple37(
simple115_vectors=np.expand_dims(obs_space_high, axis=0))[0]
# Define some useful attributes
self.action_space = [
gym.spaces.Discrete(self.env.action_space.nvec[1])
for _ in range(self.nagents)
]
self.observation_space = [
gym.spaces.Box(low=obs_space_low,
high=obs_space_high,
dtype=self.env.observation_space.dtype)
for _ in range(self.nagents)
]
self.agent_types = ['agent' for _ in range(self.nagents)]
cm = plt.cm.get_cmap('tab20')
self.agent_colors = [
np.array(cm(float(i) / float(self.nagents))[:3])
for i in range(self.nagents)
]
def _thunk():
if env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
elif 'academy' in env_id or '11' in env_id: # gfootball environments
#env = football_env.create_environment(
# env_name=env_id, stacked=('stacked' in state),
# with_checkpoints=('with_checkpoints' in reward_experiment),
# logdir=log_dir,
# enable_goal_videos=dump_scores and (seed == 0),
# enable_full_episode_videos=dump_full_episodes and (seed == 0),
# render=render and (seed == 0),
# dump_frequency=50 if render and seed == 0 else 0,
# representation='extracted')
env = football_env.create_environment(
env_name=env_id,
stacked=('stacked' in state),
rewards=reward_experiment,
logdir=log_dir,
render=render and (seed == 0),
dump_frequency=50 if render and seed == 0 else 0)
env = EpisodeRewardScoreWrapper(
env,
number_of_left_players_agent_controls=1,
number_of_right_players_agent_controls=0)
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
env.seed(seed + rank)
obs_shape = env.observation_space.shape
if str(env.__class__.__name__).find('TimeLimit') >= 0:
env = TimeLimitMask(env)
if is_atari:
if len(env.observation_space.shape) == 3:
env = wrap_deepmind(env, frame_stack=True)
elif len(env.observation_space.shape
) == 3 and 'academy' not in env_id and '11' not in env_id:
raise NotImplementedError(
"CNN models work only for atari,\n"
"please use a custom wrapper for a custom pixel input env.\n"
"See wrap_deepmind for an example.")
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
if len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = TransposeImage(env, op=[2, 0, 1])
return env
def __init__(self, version, envs, hours = 0, verbose = False, weights = None):
self.version = version
self.name = "football-ppo{}".format(version) + "-e{}"
self.path = "models/football-ppo-{}/".format(version)
self.defaults = {
"env_name": "",
"representation": "simple115",
"rewards": "scoring",
"render": False,
"write_video": False,
"dump_frequency": 1,
"extra_players": None,
"number_of_left_players_agent_controls": 1,
"number_of_right_players_agent_controls": 0,
"enable_sides_swap": False,
"parallel": 1
}
self.configs = list(map(lambda b: dict(map(lambda a: (a[0], a[1] if a[0] not in b.keys() else b[a[0]]), self.defaults.items())), envs))
self.training = SubprocVecEnv(reduce(lambda a, b: a + b, list(map(lambda config: [
lambda: football.create_environment(
env_name = config["env_name"],
representation = config["representation"],
rewards = config["rewards"],
render = config["render"],
write_video = config["write_video"],
dump_frequency = config["dump_frequency"],
extra_players = config["extra_players"],
number_of_left_players_agent_controls = config["number_of_left_players_agent_controls"],
number_of_right_players_agent_controls = config["number_of_right_players_agent_controls"],
enable_sides_swap = config["enable_sides_swap"]
) for _ in range(config["parallel"])
], self.configs)), []))
self.inputs = self.training.get_attr("observation_space")[0].shape[0]
self.outputs = self.training.get_attr("action_space")[0].n
self.verbose = verbose
if not verbose:
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
deprecation._PRINT_DEPRECATION_WARNINGS = False
logger = logging.getLogger()
logger.setLevel(logging.ERROR)
if weights == None:
self.model = PPO2(policy = MlpPolicy, env = self.training, verbose = int(self.verbose))
else:
self.model = PPO2.load(weights, env = self.training, learning_rate = 0.002)
self.experience = hours * 60
def create_single_football_env(args):
env = football_env.create_environment(
env_name=args.env_name, stacked=True,
logdir='/tmp/rllib_test',
write_goal_dumps=False, write_full_episode_dumps=False, render=args.render,
dump_frequency=0,
number_of_left_players_agent_controls=args.left_agent,
number_of_right_players_agent_controls=args.right_agent,
channel_dimensions=(42, 42))
return env
def init_env():
env = football_env.create_environment(
env_name=config["academy_scenario"],
rewards=config["scoring"],
render=config["render_mode"],
number_of_left_players_agent_controls=config["num_to_control"],
representation='simple115v2')
env.seed(seed + rank * 1000)
np.random.seed(seed + rank * 1000)
return env
def create_single_football_env(iprocess):
"""Creates gfootball environment."""
env = football_env.create_environment(
env_name=FLAGS.level, stacked=('stacked' in FLAGS.state),
rewards=FLAGS.reward_experiment,
logdir=logger.get_dir(),
write_goal_dumps=FLAGS.dump_scores and (iprocess == 0),
write_full_episode_dumps=FLAGS.dump_full_episodes and (iprocess == 0),
render=FLAGS.render and (iprocess == 0),
dump_frequency=50 if FLAGS.render and iprocess == 0 else 0)
env = monitor.Monitor(env, logger.get_dir() and os.path.join(logger.get_dir()))
return env
def __init__(self, num_agents):
self.env = football_env.create_environment(
env_name='11_vs_11_easy_stochastic', stacked=True,
logdir='logs',
write_goal_dumps=False, write_full_episode_dumps=False, render=False,
dump_frequency=0,
number_of_left_players_agent_controls=1,
number_of_right_players_agent_controls=0,rewards='scoring,checkpoints')
self.action_space = self.env.action_space
self.observation_space = self.env.observation_space
self.num_agents = 1
def get_next(self):
scenario = self.training_plan.popleft()
self.current_scenario_name = scenario
env = football_env.create_environment(
env_name=scenario,
stacked=False,
logdir='/tmp/football',
representation=self.representation,
write_goal_dumps=False,
write_full_episode_dumps=False,
render=False)
return env
def __init__(self, config):
self.mode = config["mode"]
if self.mode == 'test':
print(
"\n++++++++++++++++++++RUNNING EVALUATION MODE++++++++++++++++++++++\n"
)
# Testing mode
self.env = football_env.create_environment(
env_name='11_vs_11_easy_stochastic',
stacked=True,
logdir='/tmp/rllib_test',
write_goal_dumps=False,
write_full_episode_dumps=False,
render=False,
dump_frequency=0,
number_of_left_players_agent_controls=1,
number_of_right_players_agent_controls=0,
rewards='scoring,checkpoints')
self.num_agents = 1
else:
# Training mode
self.env = football_env.create_environment(
env_name='11_vs_11_easy_stochastic',
stacked=True,
logdir='/tmp/rllib_test',
write_goal_dumps=False,
write_full_episode_dumps=False,
render=False,
dump_frequency=0,
number_of_left_players_agent_controls=1,
number_of_right_players_agent_controls=1,
rewards='scoring,checkpoints')
self.num_agents = 2
self.observation_space = gym.spaces.Box(0,
255, [72, 96, 16],
dtype=np.float32)
self.action_space = gym.spaces.Discrete(19)
def init_env():
# (** EDITED **) Google Football Env
# Gym Env 설정 함수를 변경
env = football_env.create_environment(
env_name=config["academy_scenario"],
rewards=config["scoring"],
render=config["render_mode"],
number_of_left_players_agent_controls=config["num_to_control"],
representation='raw')
env.seed(seed + rank * 1000)
np.random.seed(seed + rank * 1000)
return env
def create_single_football_env(seed):
"""Creates gfootball environment."""
env = football_env.create_environment(
env_name=args.level,
stacked=('stacked' in args.state),
rewards=args.reward_experiment,
# logdir=logger.get_dir(),
# enable_goal_videos=args.dump_scores and (seed == 0),
# enable_full_episode_videos=args.dump_full_episodes and (seed == 0),
render=args.render and (seed == 0),
dump_frequency=50 if args.render and seed == 0 else 0)
# env = monitor.Monitor(env, logger.get_dir() and os.path.join(logger.get_dir(),
# str(seed)))
return env
def run(config):
env = football_env.create_environment(
env_name=config["academy_scenario"],
rewards=config["scoring"],
render=config["render_mode"],
number_of_left_players_agent_controls=config["num_to_control"],
representation='raw')
model = AttentionSAC.init_from_save(
"./models/football/MAAC3/run2/model.pt", True)
# (** EDITED **) Set Replay Buffer
# env.action_space, env.observation_space 의 shape를 iteration을 통해 버퍼 설정
for ep_i in range(0, config["n_episodes"], config["n_rollout_threads"]):
obs = env.reset()
obs = make_state(obs)
model.prep_rollouts(device='cpu')
for et_i in range(config["episode_length"]):
print("episode : {} | step : {}".format(ep_i, et_i), end='\r')
# rearrange observations to be per agent, and convert to torch Variable
torch_obs = [
Variable(torch.Tensor(np.vstack(obs[:, i])),
requires_grad=False) for i in range(model.nagents)
]
# get actions as torch Variables
torch_agent_actions = model.step(torch_obs, explore=True)
# convert actions to numpy arrays
agent_actions = [ac.data.numpy() for ac in torch_agent_actions]
# rearrange actions to be per environment
actions = [[ac[i] for ac in agent_actions]
for i in range(config["n_rollout_threads"])]
# Reform Actions list to fit on Football Env
# Google Football 환경은 액션 리스트 (one hot encoded)가 아닌 정수값을 받음
actions_list = [[np.argmax(b) for b in a] for a in actions]
# Step
next_obs, rewards, dones, infos = env.step(actions_list)
next_obs = make_state(next_obs)
# Prevention of divergence
# 안해주면 발산해서 학습 불가 (NaN)
rewards = rewards - 0.000001
# Reform Done Flag list
# replay buffer에 알맞도록 done 리스트 재구성
obs = next_obs
env.close()
def __init__(self, num_agents):
self.env = football_env.create_environment(
env_name='test_example_multiagent', stacked=False,
logdir=os.path.join(tempfile.gettempdir(), 'rllib_test'),
write_goal_dumps=False, write_full_episode_dumps=False, render=True,
dump_frequency=0,
number_of_left_players_agent_controls=num_agents,
channel_dimensions=(42, 42))
self.action_space = gym.spaces.Discrete(self.env.action_space.nvec[1])
self.observation_space = gym.spaces.Box(
low=self.env.observation_space.low[0],
high=self.env.observation_space.high[0],
dtype=self.env.observation_space.dtype)
self.num_agents = num_agents
def create_single_football_env(seed, level='academy_empty_goal_close'):
"""Creates gfootball environment."""
env = football_env.create_environment(
env_name=level,
stacked=('stacked' in args.state),
rewards=args.reward_experiment,
# logdir="/home/sahar/downloads",
# enable_goal_videos=args.dump_scores and (seed == 0),
# enable_full_episode_videos=args.dump_full_episodes and (seed == 0),
render=args.render and (seed == 0),
dump_frequency=50 if args.render and seed == 0 else 0)
print("Creating env:{}".format(level))
# env = monitor.Monitor(env, logger.get_dir() and os.path.join(logger.get_dir(),
# str(seed)))
return env
def create_single_football_env(seed, level):
"""Creates gfootball environment."""
env = football_env.create_environment(env_name=level,
stacked=('stacked'
in 'extracted_stacked'),
rewards='scoring',
logdir=logger.get_dir(),
enable_goal_videos=False,
enable_full_episode_videos=False,
render=False,
dump_frequency=0)
env = monitor.Monitor(
env,
logger.get_dir() and os.path.join(logger.get_dir(), str(seed)))
return env
def create_single_football_env(seed):
"""Creates gfootball environment."""
env = football_env.create_environment(
env_name=FLAGS.level,
stacked=('stacked' in FLAGS.state),
with_checkpoints=('with_checkpoints' in FLAGS.reward_experiment),
logdir=logger.get_dir(),
enable_goal_videos=FLAGS.dump_scores and (seed == 0),
enable_full_episode_videos=FLAGS.dump_full_episodes and (seed == 0),
render=FLAGS.render and (seed == 0),
dump_frequency=50 if FLAGS.render and seed == 0 else 0)
env = monitor.Monitor(
env,
logger.get_dir() and os.path.join(logger.get_dir(), str(seed)))
return env
def make_football(env_id, max_episode_steps=None):
'Modified make_atari to work with football env'
env = football_env.create_environment(env_name=env_id,
stacked=False,
representation='extracted',
write_goal_dumps=False,
write_full_episode_dumps=False,
rewards='scoring,checkpoints',
channel_dimensions=(84, 84),
render=False)
# env = NoopResetEnv(env, noop_max=30)
# env = MaxAndSkipEnv(env, skip=4) # TODO: does this make sense?
assert max_episode_steps is None
return env
def __init__(self, num_agents):
self.env = football_env.create_environment(
env_name='test_example_multiagent', stacked=False,
rewards='scoring',
#logdir='/tmp/rllib_test',
enable_goal_videos=False, enable_full_episode_videos=False, render=True,
dump_frequency=0,
number_of_left_players_agent_controls=num_agents,
channel_dimensions=(42, 42))
#self.actions_are_logits = env_config.get("actions_are_logits", False)
self.action_space = gym.spaces.Discrete(self.env.action_space.nvec[1])
self.observation_space = gym.spaces.Box(
low=self.env.observation_space.low[0],
high=self.env.observation_space.high[0],
dtype=self.env.observation_space.dtype)
self.num_agents = num_agents